Seedance uses Multi's asynchronous video task flow: submit once, save the local task_id, poll the same task, then download through the local content endpoint. It supports text-to-video, a first frame, first/last-frame transitions, multiple reference images, and reference videos.

Seedance 使用 Multi 的异步视频任务流程：只提交一次，保存本地 task_id，持续查询同一任务，成功后通过本地内容接口下载。支持文生视频、首帧图、首尾帧过渡、多图参考和参考视频。

Seedance public pricing

Seedance 公开定价

Without reference videos, billing is the selected resolution rate multiplied by generated duration. A request containing video uses Mix billing: the provider's reference-video tier is multiplied by generated duration only; the reference video's own duration is not added to billable seconds. Multi reconciles the system quote immediately after task acceptance. Failed tasks are refunded.

不带参考视频时，费用等于所选分辨率单价 × 生成时长。请求中包含 video 时使用 Mix 计费：按对应分辨率的含参考视频档位 × 生成时长计费，参考视频自身时长不计入计费秒数。任务被上游接受后，Multi 会立即按系统报价调整预扣；失败任务自动退款。

Seedance text-to-video request body:

Seedance 文生视频请求体：

{
  "model": "seedance-2.0",
  "prompt": "A paper boat drifting through a rain-washed neon street, cinematic tracking shot.",
  "duration": 4,
  "resolution": "480p",
  "aspect_ratio": "16:9"
}

Complete text-to-video submit example:

完整文生视频提交示例：

curl -X POST "https://multimodal.vibelearning.top/v1/videos" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.0",
    "prompt": "A paper boat drifting through a rain-washed neon street, cinematic tracking shot.",
    "duration": 4,
    "resolution": "480p",
    "aspect_ratio": "16:9"
  }'

Seedance first-frame image-to-video request body:

Seedance 首帧图生视频请求体：

{
  "model": "seedance-2.0",
  "prompt": "The subject turns slowly toward the camera while the fabric moves naturally in a light breeze.",
  "duration": 6,
  "resolution": "720p",
  "aspect_ratio": "9:16",
  "image_url": "https://example.com/first-frame.jpg"
}

Seedance first/last-frame transition request body:

Seedance 首尾帧过渡请求体：

{
  "model": "seedance-2.0",
  "prompt": "Create a smooth cinematic transition from the opening composition to the closing composition.",
  "duration": 8,
  "resolution": "1080p",
  "aspect_ratio": "16:9",
  "image": "https://example.com/first-frame.jpg",
  "image_tail": "https://example.com/last-frame.jpg"
}

Reference media requirements

参考素材要求

image/image_url sets the first frame, image_tail sets the last frame, images accepts up to 9 reference images, and video accepts one direct URL or an array of up to 3 direct URLs.

image/image_url 用于首帧，image_tail 用于尾帧，images 最多接受 9 张参考图；video 可传一个文件直链或最多 3 个文件直链组成的数组。

Mix request body using one strict direct-download video URL:

使用一个严格文件直链的 Mix 请求体：

{
  "model": "seedance-2.0",
  "prompt": "Use the motion and camera language from the references to create a coherent new shot.",
  "duration": 6,
  "resolution": "720p",
  "video": "https://cdn.example.com/files/reference.mp4"
}

Poll and download:

查询与下载：

curl "https://multimodal.vibelearning.top/v1/videos/{task_id}" \
  -H "Authorization: Bearer YOUR_API_KEY"

curl -L "https://multimodal.vibelearning.top/v1/videos/{task_id}/content" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -o result.mp4

Verified public routes

已验证公开路径

Base URL

请求地址

https://image1.vibelearning.top

Authentication

鉴权

Use a key on the banana group. Send it as Authorization: Bearer YOUR_API_KEY on both verified routes.

使用挂在 banana 分组上的 key。两条已验证路径都用 Authorization: Bearer YOUR_API_KEY。

Native generateContent

原生 generateContent

This is the Gemini-native Banana endpoint. Put the model name in the path. Keep generationConfig.responseModalities as TEXT plus IMAGE.

这是 Banana 的 Gemini 原生端点。模型名写在路径里。保持 generationConfig.responseModalities 为 TEXT 加 IMAGE。

curl -X POST "https://image1.vibelearning.top/v1beta/models/gemini-3.1-flash-image-preview:generateContent" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [
      {
        "role": "user",
        "parts": [
          {"text": "A single yellow banana on a white table, studio lighting, photorealistic. Square 1:1 composition."}
        ]
      }
    ],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"]
    }
  }'

Typical native success shape:

典型原生成功回包：

{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "inlineData": {
              "mimeType": "image/jpeg",
              "data": "..."
            }
          }
        ]
      }
    }
  ],
  "usageMetadata": {}
}

OpenAI-compatible chat completions

OpenAI 兼容 Chat Completions

Use this if your client already speaks OpenAI chat. The gateway converts it to native generateContent. Keep generationConfig.responseModalities as TEXT plus IMAGE. The image comes back inside the chat message, usually as a data:image/...;base64,... string.

如果你的客户端已经按 OpenAI chat 接入，用这条。网关会把它转成原生 generateContent。保持 generationConfig.responseModalities 为 TEXT 加 IMAGE。图片在聊天消息里返回，通常是 data:image/...;base64,... 字符串。

curl -X POST "https://image1.vibelearning.top/v1/chat/completions" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.1-flash-image-preview",
    "messages": [
      {
        "role": "user",
        "content": "A single yellow banana on a white table, studio lighting, photorealistic. Square 1:1 composition."
      }
    ],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"]
    },
    "stream": false
  }'

Typical success shape:

典型成功回包：

{
  "id": "chatcmpl-xxx",
  "object": "chat.completion",
  "model": "gemini-3.1-flash-image-preview",
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "data:image/jpeg;base64,..."
      }
    }
  ],
  "usage": {
    "prompt_tokens": 276,
    "completion_tokens": 1519,
    "total_tokens": 1795
  }
}

Image-to-image

图生图

Stay on the two verified Banana routes. Do not switch to /v1/images/edits.

继续走上面两条已验证路径。Banana 不要改用 /v1/images/edits。

Native: put the prompt in a text part and attach each reference as inline_data.

原生：提示词放在 text part，每张参考图作为一个 inline_data part。

curl -X POST "https://image1.vibelearning.top/v1beta/models/gemini-3.1-flash-image-preview:generateContent" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [
      {
        "role": "user",
        "parts": [
          {"text": "Keep the same subject and restyle it as a clean product poster."},
          {
            "inline_data": {
              "mime_type": "image/jpeg",
              "data": "BASE64_IMAGE"
            }
          }
        ]
      }
    ],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"]
    }
  }'

OpenAI-compatible: put the prompt in a text part and attach each reference as an image_url part.

OpenAI 兼容：提示词放在 text part，每张参考图作为一个 image_url part。

curl -X POST "https://image1.vibelearning.top/v1/chat/completions" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.1-flash-image-preview",
    "messages": [
      {
        "role": "user",
        "content": [
          {"type": "text", "text": "Keep the same subject and restyle it as a clean product poster."},
          {"type": "image_url", "image_url": {"url": "https://example.com/reference.jpg"}}
        ]
      }
    ],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"]
    },
    "stream": false
  }'

Client rules

客户端规则

• Native: parse candidates[0].content.parts. Accept both inlineData and inline_data. Decode the base64 data field.
• 原生：读取 candidates[0].content.parts。同时兼容 inlineData 和 inline_data。解码其中的 base64 data。
• OpenAI-compatible: parse choices[0].message.content. It may be a string or an array of parts. Extract every data:image URL.
• OpenAI 兼容：读取 choices[0].message.content。它可能是字符串，也可能是 parts 数组。抽出其中所有 data:image。
• Set a client read timeout of at least 60 seconds. Successful calls commonly finish in under 15 seconds, but slower peaks happen.
• 客户端读取超时至少设 60 秒。成功请求常见 15 秒内返回，但高峰时会更慢。
• Use the exact public model names above. Do not strip -preview.
• 使用上面的公开模型名，不要去掉 -preview。
• 451 Content blocked is a safety rejection, not a channel outage. Change the prompt instead of retrying the same request.
• 451 Content blocked 是内容安全拦截，不是渠道故障。应改提示词，不要原样重试。

Base URL

请求地址

https://image1.vibelearning.top/v1

Authentication

鉴权

Use a key on an image-plus series group. Send Authorization: Bearer YOUR_API_KEY. The model name is still gpt-image-2.

使用挂在 image-plus 系列分组上的 key。请求头为 Authorization: Bearer YOUR_API_KEY。模型名仍是 gpt-image-2。

Public routes

公开路由

Text-to-image example

文生图示例

curl -X POST "https://image1.vibelearning.top/v1/images/generations" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A clean sci-fi product poster, minimal composition, high detail",
    "size": "1024x1024",
    "n": 1
  }'

Async text-to-image example

异步文生图示例

curl -X POST "https://image1.vibelearning.top/v1/images/generations/tasks" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A clean sci-fi product poster, minimal composition, high detail",
    "size": "1024x1024",
    "n": 1
  }'

The create call returns immediately with a local task ID:

创建接口会立即返回本地任务 ID：

{
  "id": "task_xxx",
  "task_id": "task_xxx",
  "object": "image.generation.task",
  "status": "queued",
  "created_at": 1781803308,
  "query_endpoint": "/v1/images/generations/tasks/task_xxx"
}

Poll the task with the same Bearer token:

使用同一个 Bearer token 轮询任务：

curl -X GET "https://image1.vibelearning.top/v1/images/generations/tasks/task_xxx" \
  -H "Authorization: Bearer YOUR_API_KEY"

When finished, status becomes completed and response contains the same OpenAI-compatible image payload returned by the synchronous route, including usage when the upstream provider returns token usage.

完成后，status 会变为 completed，response 中会包含与同步接口一致的 OpenAI 兼容图片响应；如果上游返回 token 用量，也会保留 usage。

{
  "id": "task_xxx",
  "task_id": "task_xxx",
  "object": "image.generation.task",
  "status": "completed",
  "response": {
    "created": 1781803330,
    "data": [
      {
        "url": "https://image1.vibelearning.top/..."
      }
    ],
    "usage": {
      "prompt_tokens": 19,
      "completion_tokens": 734,
      "total_tokens": 753
    }
  }
}

Single-image edit example

单图图生图示例

curl -X POST "https://image1.vibelearning.top/v1/images/edits" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=请基于参考图做高精度重绘" \
  -F "size=1024x1024" \
  -F "n=1" \
  -F "image=@./reference.png"

Multi-image edit example

多图图生图示例

curl -X POST "https://image1.vibelearning.top/v1/images/edits" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=Use all reference images together to create one cohesive poster." \
  -F "size=1024x1024" \
  -F "n=1" \
  -F "image[]=@./ref1.png" \
  -F "image[]=@./ref2.png" \
  -F "image[]=@./ref3.png"

Response handling

返回处理

gpt-image-2 may return either b64_json or image URLs. Your client should support both response shapes.

gpt-image-2 可能返回 b64_json，也可能返回图片 URL。客户端应兼容这两种返回形式。

Image sync and async behavior图片同步与异步说明

Practical timeout advice: image generation can take tens of seconds, especially for high resolution or multi-image edits. Set your client read timeout high enough, and retry only after checking whether the previous request already produced a billable result.

实用超时建议：图片生成可能需要几十秒，尤其是高分辨率或多图编辑。客户端读取超时要设置得足够长；重试前请先确认上一条请求是否已经产生可计费结果。

Image output formats图片输出格式

Base URL is https://image1.vibelearning.top/v1. Use a 原生 4k key as Authorization: Bearer YOUR_API_KEY. Do not use an image-plus key here.

请求地址为 https://image1.vibelearning.top/v1。使用 原生 4k 分组的 key，请求头 Authorization: Bearer YOUR_API_KEY。不要用 image-plus 系列的 key。

Sizes

尺寸表

Pass size as widthxheight. Do not put 4k into quality. Check decoded pixels, not only the echoed size.

用 widthxheight 传 size。不要把 4k 写进 quality。验收看解码像素，不要只看回显 size。

quality is low / medium / high. detail_level is 1–5 and overrides quality. Neither changes the ¥0.10 unit price. 4K calls need a 90s read timeout.

quality 取 low / medium / high。detail_level 为 1–5，会覆盖 quality。两者都不改 ¥0.10 单价。4K 读取超时至少 90 秒。

Examples

示例

curl -X POST "https://image1.vibelearning.top/v1/images/generations" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-image-2","prompt":"a cinematic mountain sunrise","size":"3840x2160","quality":"high"}'

curl -X POST "https://image1.vibelearning.top/v1/images/edits" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=turn this photo into watercolor style" \
  -F "size=2496x1664" \
  -F "image=@./input.png"

curl -X POST "https://image1.vibelearning.top/v1/images/edits/tasks" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=blend these images into one scene" \
  -F "image[]=@./reference-a.png" \
  -F "image[]=@./reference-b.png"

Edits are whole-image redraws, 1–6 references, no mask. Async edits are open; poll GET /v1/images/edits/tasks/{task_id} with the same token and read response.data when completed.

图生图是整体重绘，1–6 张参考图，不支持 mask。异步图生图已开放；用同一个 token 轮询 GET /v1/images/edits/tasks/{task_id}，completed 后读 response.data。